Accurate multi-view depth reconstruction with occlusions handling

Abstract

We present an offline method for stereo matching using a large number of views. Our method is based on occlusions detection. It is composed of two steps, one global and one local. In the first step we formulate an energy function that handles data, occlusions, and smooth terms through a global graph-cuts optimization. In our second step we introduce a local cost that handles occlusions from the first step in order to refine the result. This cost takes advantage of both the multi-view aspect and the occlusions. The experimental results show how our algorithm joins the advantages of both global and local methods, and how much it is accurate on boundaries detection and on details.